40 lines
1.5 KiB
Markdown
40 lines
1.5 KiB
Markdown
---
|
||
title: "鲁棒性认证 (Robustness Certification)"
|
||
created: 2026-07-04
|
||
updated: 2026-07-04
|
||
type: concept
|
||
tags: [robustness, certification, verification, safety]
|
||
sources: []
|
||
---
|
||
|
||
# 鲁棒性认证 (Robustness Certification)
|
||
|
||
鲁棒性认证的目标是给出模型预测在一组允许输入变换下保持不变的保证。给定变换 $\gamma$ 和强度参数 $\varphi$，认证目标是确定预测保持不变的 $\varphi$ 范围。
|
||
|
||
## 方法分类
|
||
|
||
### 概率方法
|
||
- **随机平滑** ([[randomized-smoothing]])：通过噪声注入给出置信度意义下的认证
|
||
- PixelDP：差分隐私视角的鲁棒性保证
|
||
|
||
### 确定性不完整验证
|
||
- DeepPoly ([[deep-poly]])、CROWN ([[crown-verifier]])、PRIMA：凸松弛/抽象解释，给出 sound 但保守的保证
|
||
|
||
### 完备验证
|
||
- ReluVal、$\beta$-CROWN (branch-and-bound)、MN-BaB：精确性优先
|
||
|
||
### 语义变换认证
|
||
- ExactLine ([[exactline]])：沿图像线性插值路径的完备认证
|
||
- ApproxLine、GCERT：生成模型潜空间中的语义变化认证
|
||
|
||
## 与像素扰动认证的区别
|
||
|
||
传统认证多关注 $L_p$ ball 内的像素扰动，难以表达语义层面的变化（如"变圆""变暗""换风格"）。语义鲁棒性认证将变换建模推进到语义层。
|
||
|
||
## 参考
|
||
|
||
- [[semantic-robustness-certification-vlm-2026|Semantic Robustness Certification (ICML 2026)]]
|
||
- [[semantic-robustness-certification|语义鲁棒性认证]]
|
||
- [[adversarial-robustness|对抗鲁棒性]]
|
||
- [[distribution-shift|分布偏移]]
|